> One successful deep-learning approach involves modeling behavioral tasks. Researchers build an ANN and optimize it to solve a task analogous to the one studied in animals. They then compare the internals of the trained ANN with the biological neural recordings, typically smoothed spike trains of a population of neurons. If there are quantitative similarities, researchers can then attempt to reverse-engineer the ANN in order to develop a mechanistic explanation for how the ANN solves the task. The insights found in the ANN can lead to testable hypotheses for how the biological network implements the behavior.
So, IIUC, the approach involves (motivated) guesswork over the space of practically implementable artificial neural network architectures (which is a strongly biased subset of the set of all actual neural networks), to mimic behavior that is presumably fully captured by biological neural networks.
While it is tremendously exciting to try something new and find surprises, why would I expect this to be a fruitful systematic approach to neuroscience? I.E. choose a particular behavior, find an artificial network solving it, and then "understand" why it works biologically.
Basically comes down to two noob questions:
1. Why should we expect the search over architectures of artificial network architectures to be efficient/fruitful?
2. How do we know that the brain might not use a completely different architecture since computationally implementable NNs have many artificial limitations.
I would appreciate any insights and expert comments. Thanks!
> How do we know that the brain might not use a completely different architecture since computationally implementable NNs have many artificial limitations.
not just ANNs have limitations, but biological networks as well. Degeneracy is biology's version of "more than one way to do it" and it's present everywhere; in other words, different brain networks can implement same function. One example of this is that when you write your signature with your foot, it shows similar characteristics as when you write with your hand.
Degeneracy and complexity in biological systems is a good read in this direction:
> How do we know that the brain might not use a completely different architecture since computationally implementable NNs have many artificial limitations.
I'd like to add this question: have we found the natural equivalent of "backpropagation" yet?
There's a somewhat similar mechanism which underlies all learning. It's "what fires together, wires together", meaning that neural connections with correlated activity patterns tend to get strengthened over time.
While that mechanism operates on what at first seems to be the forward pass, the neurosystem seems to often operate in loops, with "useful" signals continuing to go around for a while to provide the learning signal.
One of the proposals for a biologically realistic alternative to backpropagation is feedback alignment: http://www.cs.toronto.edu/~tingwuwang/2546.pdf (randomly selected representative slides from Google search. See reference therein)
As an engineer in a systems neuroscience lab, I'm likely biased, but this is a misguided point of view: sufficiently flexible models for data have been around for a long time, but they're not useful because they provide no insight into the mechanisms at work. A DL model which might perform 100% still requires study itself to produce anything with explanatory value.
edit to add, we are in fact doing projects with DL on real data, and it's extremely difficult to reverse-engineer the trained DL model to figure out how/why it does what it does: the parameters are simply uninterpretable without a good theory or mechanism, and this is precisely what is sought.
What if the models just can't be distilled beyond a certain point down to an explanation in English words that makes sense to you and has "explanatory power"?
Clearly reducing the problem from the brain to an ANN is valuable because if we want to predict build or fix the brain, having approximations to pieces of it as an ANN let us get closer to doing that in the same way that more compact models or explanations help us get closer to doing those things.
At the moment, the way they are trained is not a good theory. But that is the compact human interpretable way of thinking of these models. It seems like if we keep iterating on this then we could arrive at a compact description of the neural network which is its learning rules, architecture and environment. Why is it important to have a compact explanation of the trained resulting model if the learning rule, architecture and data are a fairly compact description?
It seems like for vision there are a few simple theories of learning:
having layers of nonlinearities
weight sharing across space
and some way of doing credit assignment on the loss from a visual task
Which taken together are enough to explain a large amount of the explainable variance in the neural data. I agree that the models could get more biologically realistic in the way they learn, but I disagree that it's important to explain how the learned model functions in a compact way, since there may be no such explanation better than the one based on learning.
Major takeaway : We are too bored to bother with really understanding stuff from first principles, let the computer do my work and I'll write a paper that I made the computer do my work.
"We are all familiar with the standard paradigm in systems neuroscience"
I don't know why but that opening cracked me up. I'm sure everyone who frequents the site is actually familiar, but reading that opener felt like being back in high school and being asked about the book you didn't read.
IMHO, they could, maybe should, with some promise start with what are likely relatively low level brain tasks say, among the first things a brain does with an input via touch, temperature, sound, sight, movement.
In a computer analogy, pay attention to the processor, main memory, Ethernet connection, PCI bus, etc.
IMHO, for the usual meanings of behavior, e.g., a kitty cat chases a mouse, a kitty cat jumps successfully from a kitchen stool to the kitchen counter top that has the offered food and water, a kitty cat follows its owner from room to room, etc. are at too high a level.
With a computer analogy, such behavior doesn't really directly involve the processor instruction set, the word length in main memory, communications over Ethernet or USB but involves software.
As we know well, given inputs and outputs of some software, even millions of pairs of inputs and outputs, tough to infer the source code or even the programming language used, C++, C, C#, Lisp, Fortran, etc.
Still, we know some things about software, e.g., the programs commonly use Do-While, If-Then-Else, Call-Return, raise exceptional condition. So, there's a microscopically thin chance that the coveted theory of neuro-science could have the coveted theories with testable hypotheses at roughly that level.
For more about behavior, IMHO have to accept real cognition, actually thinking, e.g., considering scenarios and possibilities for responding. We can guess that the work of this cognition will as software in a computer use relevant accumulated data, some usually crude causal models, and some simple deductive logic.
Can the line of attack in the OP address such cognition? Maybe, but here being closely reductionist standing on solid theories of the lower level functioning is likely asking for too much and maybe, for some progress, not necessary. E.g., commonly parts of psychology study cognition while largely ignoring the lower level neuro-science mechanisms.
Part of the trouble is that the processor, main memory and such are not really discrete. The brain and nervous system is far more interconnected than that. So what's happening isn't that the something in a discrete memory changes, but the circuits themselves change.
What you are suggesting is, IIRC, far more high level than where the field has been. What the author is suggesting is that they now have the ability to gain larger samples than from a single neuron, so they're suggesting use cases for the new data. This is somewhat akin to examining what is happening at several logic gates within a microprocessor, probably quite a bit lower level than what you're suggesting.
Have to take analogies with current computers with a large shovel full of salt.
The OP says more than once that they are trying to understand "behavior". To me that is like the kitty cat chasing a mouse, and in that case I suggested starting at a level lower than that. Sure, maybe as you suggest that level is higher than neuro-science has achieved so far, but I was responding to the OP.
A few days ago I did conjecture that building models of brain activity might give some clues to how the brain actually works, maybe.
If memory in brains involved rewiring, than maybe could treat that memory just as memory, although based on rewiring. In that case, maybe would want to study the wiring and, if some memory changed, the new wiring and, thus, see how memory writing is implemented by neuron rewiring. Then, maybe, the more times that memory is read and the length of time the reading is active determines how strong the connections are in the rewiring and, thus, how persistent the memory is. Maybe.
They take basic neuronal growth "laws" inferred by Ramon y Cajal just by looking at a crapload of stained neurons in the late 19th to early 20th century and apply modern computational techniques to grow realistic neurons based on randomly placed "growth signals" and an extremely simple rule.
Of course the next step is to grow groups of these "neurons" together with non-randomly placed signals based on some sort of input and allowing one to input on the other. I haven't followed up on it for a few years so maybe they've gotten into that.
To me this looks like successful science in action. Something that looks very complex (dendritic arbors) turns out to be explainable by a very simple to understand process/rule/principle.
> low-cost recording technologies that are easy to use, such as the Neuropixels probes
I'm going to withhold my belief that it's going to go bonkers due to this technology. Things go bonkers when they're useful to people and money is behind it. Neuropixel probes require the scientist to shove silicon into a brain to measure it. So, not very useful for people. If we figured out cheap low temperature superconductors, and everyone courd afford an fMRI machine at home, that would be bonkers.
With that said, I think brain research is ready to blossom. We now have proven models for neural nets, and the compute capacity is getting there. The problem is that it is still too expensive (cost, signal/noise, physical discomfort) to reliably read (complex) brains.
I don't know anything about neuropixels (added that Nature paper to my queue), so don't know if they live up to the hype. That said, the standard tool in many electrophysiology experiments are still tetrodes and variants (https://en.wikipedia.org/wiki/Tetrode_(biology)), which can at best record a few to several cells at a time, within a small radius of the recording tip. Simultaneously recording from hundreds to thousands of cells in multiple brain areas would probably yield quite a bit of valuable data.
Also, non-invasive techniques like fMRI (https://en.wikipedia.org/wiki/Functional_magnetic_resonance_...) tend to have relatively poor spatial resolution, and track signals that correlate with neural activity (e.g., blood flow) rather than actual electrical activity. Since nearby neurons do not always encode the same information or perform the same function, a spatially-averaged signal isn't going to be as informative.
So, IIUC, the approach involves (motivated) guesswork over the space of practically implementable artificial neural network architectures (which is a strongly biased subset of the set of all actual neural networks), to mimic behavior that is presumably fully captured by biological neural networks.
While it is tremendously exciting to try something new and find surprises, why would I expect this to be a fruitful systematic approach to neuroscience? I.E. choose a particular behavior, find an artificial network solving it, and then "understand" why it works biologically.
Basically comes down to two noob questions:
1. Why should we expect the search over architectures of artificial network architectures to be efficient/fruitful?
2. How do we know that the brain might not use a completely different architecture since computationally implementable NNs have many artificial limitations.
I would appreciate any insights and expert comments. Thanks!